Chapter 17
The Generalized Monte Carlo Particle Filter
All of the SIS class of particle filters discussed in the previous chapter are based on (16.12), which defines the recursive update of the importance weights. The combination of this recursive importance weight update with the resample and move steps provide the framework for powerful nonlinear non-Gaussian estimation methods dependent only on the choice of importance density and likelihood function. However, there are many disadvantages in using the SIS particle filters:
- The BPF does not use the latest measurement during importance sampling leaving it susceptible to increased variance and instability for many applications.
- All SIS particle filters require the resample and move steps that tend to increase the estimated state error covariances. They also add a significant computational burden to the estimation procedure, since they cannot be parallelized.
- The SIS APF and its alternatives may require the execution of a Gaussian Kalman filter for every particle during every time step, increasing the computational burden even further.
A second class of particle filters, the generalized monte carlo (GMC) particle filters, have been developed that do not use a recursive weight update, but instead compute new weights with every sequential update of the filter. An example of this more general particle filter class include the Gaussian particle filter (GPF) first proposed by Kotechha and Djuri [1,2].
